Re: Compiler optimization?

[Jim Wilson <wilson at tuliptree dot org>]

Joe Buck wrote:
> On Tue, Jun 10, 2003 at 05:41:13PM -0700, Chirag Wighe wrote:
> > I am an embedded developer and I am trying to understand the endian issues 
> > with respect to compilers and how they access structure fields.

> > Is it possible for an optimizing compiler to combine the two half word 
> > writes into a single word write.

Gcc isn't very good about optimizing stores, but it will optimize the 
loads.  If you load one of the shorts, gcc will load an int instead.  If 
you load both of the shorts near each other, one of the int loads will 
be optimized away.  (This is assuming a 32-bit target.)  This will fail 
in your case.  You can prevent that by using volatile as Joe Buck mentioned.

> Obviously the 32-bit constant that would be stored depends on the
> endian-ness of the target.

Some systems allow one to specify data endianness separately from cpu 
endianness.  I think this is what he is referring to.  This can be 
useful when moving data between two systems with different 
cpu-endianness.  Gcc doesn't directly support this.  There have been 
occasional requests for an endianness attribute for data, but I am not 
convinced that it is a good idea.

See for instance the IA-64 assembler which has .lsb and .msb directives 
which specify endianness of the current section only.  This allows you 
to assemble a file that has both big and little endian data.

Jim